Inkling Is Mira Murati’s First Model. Here’s How to Actually Use It

Thinking Machines Lab, founded by OpenAI's former chief technology officer Mira Murati and sitting on two billion dollars of funding, shipped its first model Inkling on July 15. Inkling is a 975-billion-parameter open-weights mixture-of-experts transformer with a million-token context window, and the company stated in its announcement that it isn't the strongest model available, open or closed, a rare admission that reflects a strategy focused on customizability over benchmark supremacy.

Thinking Machines Lab, founded by OpenAI’s former chief technology officer and sitting on two billion dollars of funding, shipped its first model on July 15. Inkling is a 975-billion-parameter open-weights system with a million-token context window, and in its own announcement the company states plainly that it isn’t the strongest model available, open or closed. Saying that out loud on launch day is close to unheard of, and it isn’t modesty. It’s the entire strategy, a bet that the future belongs to models you can make your own rather than models you rent. Here’s what Inkling actually is, and why the honest positioning is the most interesting thing about it. Model launches follow a script. The company claims the top spot on a set of benchmarks, publishes a chart where its bars are tallest, and lets the coverage write itself. Thinking Machines Lab did something else on July 15. It released Inkling, its first model built from scratch, and stated in the announcement that Inkling isn’t the strongest overall model available today, open or closed. That sentence is genuinely unusual. Companies don’t typically spend a year and enormous resources building a model and then lead with what it doesn’t do. But read it as strategy rather than humility and it snaps into focus. Thinking Machines isn’t trying to win the leaderboard. It’s making a specific and contrarian bet about where AI value is heading, and the model it shipped is designed to serve that bet rather than a benchmark chart. Understanding what the company is actually claiming, and what it’s deliberately not claiming, is more instructive than any score. Start with the specifications, because they matter for what follows. Inkling is a mixture-of-experts transformer with 975 billion total parameters and 41 billion active on any given forward pass. Mixture-of-experts means the model contains many specialized sub-networks and routes each token through only a fraction of them, so you get the knowledge capacity of a very large model while paying the compute cost of a much smaller one. Its context window runs to a million tokens, and it was pretrained on 45 trillion tokens spanning text, images, audio, and video. It reasons natively over text, images, and audio, not through bolted-on adapters but as part of how it was trained. The weights are fully open and published on Hugging Face, which is the part that matters most to the strategy. Anyone can download the model, inspect it, run it on their own hardware, and modify it. Note the honest boundary though, open weights isn’t the same as open source. You get the trained parameters, not the training data or the pipeline that produced them, so you can use and adapt the model without being able to reproduce it. That distinction gets blurred constantly in coverage and it’s worth keeping straight. There are a few genuinely interesting technical choices inside. The model has a controllable thinking effort dial, letting you trade inference cost against reasoning depth on a per-request basis, which is a practical acknowledgment that not every query deserves the same amount of computation. Architecturally, it uses a learned, input-dependent relative-position scheme rather than the rotary position embeddings that have become standard, with the company claiming better extrapolation to long sequences, and the majority of its 66 decoder layers use local attention over a small window rather than attending globally. Those are the kinds of decisions that are interesting to people who care about how these systems are built, and they suggest a team willing to depart from the default recipe. A smaller sibling, Inkling-Small, was previewed alongside it with 12 billion active parameters, though as of shortly after launch its full weights had not been released. Now the strategy, because the model only makes sense in light of it. The dominant way to use AI today is to rent it. You call an API belonging to one of a handful of labs, you pay per token, and you get whatever the current version of their general-purpose model does. It works, it’s easy, and it means the model is a service you access rather than an asset you hold. Thinking Machines is betting that a meaningful share of organizations will eventually want the opposite, a model they own, that runs where they choose, and that has been shaped specifically for their work. That’s why Inkling is designed as a base for customization rather than a finished product. The company pairs it with Tinker, its fine-tuning platform, and positions Inkling as the flagship model to build on there. The pitch is that a broad, balanced, well-rounded foundation model that you then specialize on your own data can beat a stronger general model at your particular task, while costing less to run. Breadth over peak capability, deliberately, because a base model for fine-tuning wants to be flexible and adaptable more than it wants to top any single chart. This reframes the “not the strongest” admission entirely. If you’re selling a finished assistant, being second-best is a problem. If you’re selling a foundation that customers will specialize themselves, being broad, adaptable, and open matters more than winning benchmarks that measure general performance out of the box. The company is claiming its model is the best open-weights base for customization, which is a different claim from being the best model, and it’s a claim you evaluate on your own data rather than on a public leaderboard. It’s also a direct answer to a trend that has been reshaping the industry, the steady commoditization of raw model access. When the general-purpose model becomes a cheap, interchangeable utility, the differentiated thing isn’t access to intelligence but intelligence shaped around your specific problem and data. Thinking Machines is positioning itself on that side of the line, selling adaptability rather than raw capability. Whether enterprises actually want that badly enough to do the work of fine-tuning is the open question the whole company rests on. A few things temper the picture, and they matter for anyone considering building on this. The benchmark numbers published at launch come from the company itself. Independent third-party evaluations had not been published in the days after release, which means the performance claims, including the modest ones, are self-reported. That’s normal for a launch, and Thinking Machines deserves credit for framing its own results conservatively rather than cherry-picking, but the appropriate stance is still to wait for outside testing before treating any number as settled. Running Inkling yourself isn’t a casual undertaking. The published minimums start around 600 gigabytes of aggregate video memory for the quantized checkpoint, and roughly two terabytes for the full-precision version, before serving overhead. This isn’t a model you run on a workstation, or even on a modest cluster. The open weights are genuinely open, but exercising that openness requires serious hardware, which is why the model is also served through a range of hosted providers for anyone who wants to use it without owning the infrastructure. Open weights lower the ceiling on lock-in, they don’t eliminate the cost of independence. And the company itself has had a turbulent stretch. Thinking Machines raised an enormous seed round at a high valuation before shipping anything, and in early 2026 it lost several senior people, including departures back to OpenAI and to another major lab, with reporting suggesting a subsequent funding round stalled. None of that says anything about whether Inkling is a good model, but it’s relevant context for anyone weighing how much to build on a young company’s platform. Talent concentration is both the moat and the vulnerability in this field. If the thesis interests you, here is the practical path from reading about Inkling to running it, in rough order of effort. The lowest-friction option is to call it through a hosted provider. Inkling is served through several inference platforms, including Together, Fireworks, Modal, Databricks, and Baseten, so you can use it the same way you would any other API model without owning any hardware. Thinking Machines also exposes it through its own Tinker service using a standard chat-completions shape, which means if you have code that already talks to an OpenAI-compatible endpoint, pointing it at Inkling is close to a two-line change. curl "https://tinker.thinkingmachines.dev/services/tinker-prod/oai/api/v1/chat/completions" \ -H "Authorization: Bearer $TINKER API KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "thinkingmachines/Inkling", "messages": {"role": "user", "content": "Explain mixture-of-experts in two sentences."} , "stream": false }' The thing worth experimenting with here is the thinking effort control, which lets you dial reasoning depth against cost per request. The useful test is to run the same realistic task at a low effort setting and a high one and see where the quality actually changes for your work, because that dial is where the cost story lives. Many tasks will not need the expensive end. The second option is downloading the weights. They are published on Hugging Face, in both the original checkpoint and a quantized version aimed at current-generation hardware. This is the option that gives you true independence, the model on your own machines with nothing leaving your network. Be realistic about the requirement though. Published minimums start around 600 gigabytes of aggregate video memory for the quantized checkpoint and roughly two terabytes for full precision, before serving overhead, so this is a data-center-class undertaking rather than something for a workstation. For most teams the honest sequence is to prototype against a hosted endpoint and only consider self-hosting once the economics or the privacy requirements clearly justify the infrastructure. The third option, and the one the entire strategy is built around, is fine-tuning. Inkling is available for customization on Tinker with context length options for training, which is where the “make it your own” claim gets tested. The meaningful experiment is narrow rather than broad, pick one workflow where you have real data and a clear measure of success, fine-tune on that slice, and compare the specialized result against a strong general model on the same task. That comparison is the only thing that settles whether the customization thesis works for you, and it is a far better use of an afternoon than reading anyone’s benchmark chart, including this article’s summary of one. Strip away the specifics and Inkling is interesting for what it represents rather than what it scores. The first thing it demonstrates is that a well-funded new lab can, in roughly a year, train a large, capable, multimodal model from scratch and give the weights away. That’s a statement about how much the barrier to building frontier-adjacent models has fallen, and it adds another serious open-weights option to a field where the strongest models have mostly been closed. More open models with real capability is straightforwardly good for developers, researchers, and anyone who wants the option of running a model on their own terms. The second is the honesty itself, which deserves more notice than it’ll get. A company that tells you upfront what its model is not good at is giving you something more useful than a chart, it’s giving you an accurate basis for deciding whether the model fits your problem. The industry norm of claiming the top of every benchmark has made launch-day numbers nearly uninformative. A lab that says plainly “this is not the strongest, here is what it is actually for” is behaving the way you would want a vendor to behave, and if that framing gets rewarded, it’s good for everyone evaluating these tools. The third is the strategic question the whole thing poses. Is the future of applied AI renting the smartest general model, or owning a specialized one. Most of the industry has bet on the first. Thinking Machines has bet, publicly and expensively, on the second. It’s genuinely unclear which is right, and it may well be both for different customers. But it’s a real bet with a real product behind it, and the answer will shape where value in this industry ends up sitting. For now, the practical advice is the one the company implicitly gives. Don’t take the benchmarks, theirs or anyone’s, as the answer. If the customization thesis appeals to you, the honest test is to fine-tune it on a narrow slice of your own work and see whether a specialized version of a good model beats a general version of a great one. That’s the claim, and it’s the only test that settles it. Specifications, benchmark figures, and pricing here come from Thinking Machines Lab’s own July 2026 announcement and materials, with independent evaluation not yet available at the time of writing. Model availability, pricing, and platform details change quickly, so check current documentation before building on any of it. Inkling Is Mira Murati’s First Model. Here’s How to Actually Use It https://pub.towardsai.net/inkling-is-mira-muratis-first-model-here-s-how-to-actually-use-it-268f7cc1bfdd was originally published in Towards AI https://pub.towardsai.net on Medium, where people are continuing the conversation by highlighting and responding to this story.